A Hybrid Ensemble Framework for Robust Deepfake Image Detection Using Convolutional and Transformer Architectures

International Journal of Research in Engineering & Science ISSN:(P) 2572-4274 (O) 2572-4304 Available online on http://rspublication.com/IJRES/IJRE.html volume 9 Number 6, 2025 DOI: 10.5281/zenodo.17935569 ©2025 RS Publication, [email protected] 206 Original Article A Hybrid Ensemble Framework for Robust Deepfake Image Detection Using Convolutional and Transformer Architectures Vijaylaxmi Inamdar Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management Bangalore, India, [email protected] Jatavallabhula Sarat Anirudh Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management Bangalore, India, [email protected] Kalash Verma Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management Bangalore, India, [email protected] Manish S Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management Bangalore, India, [email protected] Komal Raj Computer Science and Engineering, Dayananda Sagar Academy of Technology and Management Bangalore, India, [email protected] International Journal of Research in Engineering & Science Available online on http://rspublication.com/IJRES/IJRE.html ISSN:(P) 2572-4274 (O) 2572-4304 ARTICLE INFO ABSTRACT ©2025 RS Publication Paper ID: : IJRES693EC48339F38 Published: 2025-12-15 DOI: https://dx.doi.org/1 0.5281/zenodo.179355 69 Page No: 206-212 I see that the fast spread of media called deepfakes comes from the powerful abilities of modern Generative AI. From diffusion models, to rendering and GAN-based synthesis the tools that make - realistic forgeries are now easy to get. Deepfakes can help industries. Bad actors can use deepfakes, for misinformation campaigns, identity theft, cyber-extortion and political manipulation. Deepfakes are a danger to trust and online safety. Detecting the manipulations is getting harder. As generative models improve the visual artifacts which traditional detectors relied on— blurring, lighting—are disappearing. I compare four learning models, for deepfake image detection: the basic CNN, the EfficientNet-B0 that focuses on texture the Xception-41 model and the Vision Transformer called ViT-Tiny that uses context. I see a problem, with the single-model detectors: single-model detectors often overfit to a certain type of manipulation and single-model detectors fail when single-model detectors meet new kinds of fakes. To solve this problem, we propose a hybrid stacking that mixes the texture analysis of Convolutional Neural Networks with the long-range structural reasoning of Transformers. We test the hybrid stacking, on the Labeled Deepfake Image Collection (LDIC)—the collection has 18,124 images made by ten techniques. The stacking ensemble gets an accuracy of 99.93% an F1-score of 0.9995 and a ROC-AUC of 1.0000. In my view these results show that mixing the two architectures is a way to make digital forensics more scalable and robust. Index Terms—Deepfake Detection, Ensemble Learning, Vision Transformer, Efficient Net, Exceptions, Digital Forensics, Misinformation, Synthetic Media Security Cite This Paper: Jatavallabhula Sarat Anirudh, Vijaylaxmi Inamdar, Kalash Verma, Manish S and Komal Raj (2025). "A Hybrid Ensemble Framework for Robust Deepfake Image Detection Using Convolutional and Transformer Architectures". INTERNATIONAL JOURNAL OF RESEARCH IN ENGINEERING & SCIENCE (IJRES), vol. 15, no. 6, 2025, pp. 206-212, DOI: https://dx.doi.org/10.5281/zenodo.17935569 [email protected] [email protected] [email protected] [email protected] [email protected] I compare four learning models, for deepfake image detection: the basic CNN, the EfficientNet-B0 that focuses on texture the Xception-41 model and the Vision Transformer called ViT-Tiny that uses context. I see a problem, with the single-model detectors: single-model detectors often overfit to a certain type of manipulation and single-model detectors fail when single-model detectors meet new kinds of fakes. To solve this problem we propose a hybrid stacking that mixes the texture analysis of Convolutional Neural Networks with the long-range structural reasoning of Transformers. We test the hybrid stacking, on the Labeled Deepfake Image Collection (LDIC)—the collection has 18,124 images made by ten techniques. The stacking ensemble gets an accuracy of 99.93% an F1-score of 0.9995 and a ROC-AUC of 1.0000. In my view these results show that mixing the two architectures is a way to make digital forensics more scalable and robust. Recent work, in image synthesis has made the fake facial images look more real. Generative Adversarial Networks (GANs) and diffusion models can produce the pictures that look like the photos. Generative Adversarial Networks (GANs) and diffusion models can help with the media creation and the visual effects. Generative Adversarial Networks (GANs) and diffusion models also bring risks such as the misinformation, the identity theft and the digital fraud. Because of the risks the need, for automated deepfake image detection grows. Early deepfake detection relied on spotting artifacts or biological mismatches in the generated media. As generative models improved the artifacts became more subtle. Vanished entirely. Modern deepfake images keep the texture and the facial geometry so human eyes often miss the deepfake images. I have watched deepfake detection move from the checks used by deepfake detection to the use of data driven deep learning techniques. Data driven deep learning techniques learn features that go beyond what human eyes can see. Convolutional Neural Networks (CNNs) are used a lot, for deepfake detection because the CNNs catch small texture changes that appear when images are made and enlarged. I have seen the CNNs in action. The XceptionNet and the EfficientNet designs have shown results, on test sets. I have International Journal of Research in Engineering & Science, ISSN:P-2572-4274, O-2572-4304 volume 9 Issue 6 2025 Available online on http://rspublication.com/IJRES/IJRE.html DOI:10.5281/zenodo.17935569 ____________________________________________________________________________________________________________________ ©2025 RS Publication, [email protected] _________________________________________________________________________________________________________________________________________________ 207 A Hybrid Ensemble Framework for Robust Deepfake Image Detection Using Convolutional and Transformer Architectures Vijaylaxmi Inamdar Computer Science and Engineering Dayananda Sagar Academy of Technology and Management Bangalore, India Jatavallabhula Sarat Anirudh Computer Science and Engineering Dayananda Sagar Academy of Technology and Management Bangalore, India Kalash Verma Computer Science and Engineering Dayananda Sagar Academy of Technology and Management Bangalore, India Manish S Computer Science and Engineering Dayananda Sagar Academy of Technology and Management Bangalore, India Komal Raj Computer Science and Engineering Dayananda Sagar Academy of Technology and Management Bangalore, India I. INTRODUCTION Abstract—I see that the fast spread of media called deepfakes comes from the powerful abilities of modern Generative AI. From diffusion models, to rendering and GAN-based synthesis the tools that make -realistic forgeries are now easy to get. Deepfakes can help industries. Bad actors can use deepfakes, for misinformation campaigns, identity theft, cyber-extortion and political manipulation. Deepfakes are a danger to trust and online safety. Detecting the manipulations is getting harder. As generative models improve the visual artifacts which traditional detectors relied on— blurring, lighting—are disappearing. Index Terms —Deepfake Detection, Ensemble Learning, Vision Transformer, EfficientNet, Xception, Digital Forensics, Misinformation, Synthetic Media Security found that the CNN-based detectors often do not work well on kinds of fakes. The reason is that the CNN-based detectors look at view windows and the CNN-based detectors pick up artifacts that belong to a specific generator. Because of that the CNN-based detectors fail on images made by synthesis methods. I have started looking at Vision Transformers (ViTs) as a way to spot deepfakes. Vision Transformers (ViTs) are powerful. Vision Transformers (ViTs) use self attention to learn range connections and the overall picture layout. Vision Transformers (ViTs) can find meaning mismatches that convolutional filters often miss. With those strengths Vision Transformers (ViTs) need data sets and more compute power. In this study we look at whether mixing the networks and the transformer networks can make the system stronger to deepfake generation methods. We propose a hybrid stacking that combines the predictions, from the EfficientNet model, the Xception model and the Vision Transformer model. We evaluate the ensemble, on the Labeled Deepfake Image Collection (LDIC). The LDIC includes images that were generated by synthesis pipelines. The experimental results show that the ensemble outperforms each model. The ensemble also shows generalization to deepfake sources. I found the results convincing because the ensemble clearly beats each model. The fight to separate the media from the forgeries has defined the digital forensics for almost ten years. The cat and mouse dynamic drives the field. When the detection method uncovers a flaw, in the model the creators patch the flaw. The creators patch the flaw. The researchers must create detection paradigms. The section follows the change from the checks, to the complex multi-modal systems that the field needs today. A. The Era of ”Sanity Checks” (Biological Signals) In the days of deepfake technology (circa 2017-2018) I could spot a fake without using neural networks; I only needed common sense. The early Generative Adversarial Networks (GANs) looked good on paper. The early Generative Adversarial Networks (GANs) ignored biology. It was simple. The early Generative Adversarial Networks (GANs) treated faces as textures of, as 3D biological structures. Li et al. handcrafted published a study that identified the blinking anomaly. The study showed that deepfake models were trained on scraped datasets of celebrities. The celebrities are usually photographed with their eyes open. Because of that the deepfake models did not learn that people need to blink. The deepfake models then produced videos where the subjects stared with a zombie, like intensity. The detection method that used the blinking anomaly was simple, explainable and very effective. I think the detection method was short lived. I saw that within weeks of the paper’s release deepfake developers added blinking samples, to the training loops. The artifact then vanished. The artifact disappeared before anyone could notice the artifact. Similar work looked at head geometry. Yang et al. HeadPose saw that when a synthetic face rotated the facial landmarks (nose, eyes, mouth) moved separately from the head pose. The movement created a sliding effect that broke physics. I have seen that the biological sanity checks built the base, for the field. I have seen that the biological sanity checks are now mostly obsolete because the modern neural rendering engines force the consistency. B. The Deep Learning Pivot (CNNs) I saw that when visual artifacts became hard to see for the eye the forensic community changed direction. If we could not see the fake maybe a computer could find the math trace of how the fake was made. That change started the rise of Convolutional Neural Networks (CNNs). The core assumption was that the upsampling operations used to make an image must leave a leftover mark—a fingerprint. Afchar et al. Proposed MesoNet [10]. I noticed that MesoNet is a network that looks for these signatures at a mesoscopic level. That means MesoNet examines mid level textures of high level semantics. MesoNet runs fast. Mesonet breaks easily when it sees high resolution inputs. Later the used model changed to **XceptionNet**. R¨ ossler et al. Changed **XceptionNet**, for work [13]. **XceptionNet** uses convolution layers on the FaceForensics++ benchmark. **XceptionNet** finds the compression marks left by face swapping algorithms. **overfitting** soon showed up as a problem. The CNN learned the pixel patterns of a StyleGAN image. When the CNN saw a Diffusion-generated image the CNN failed badly. The network was not learning the idea of ”fakeness”. The network was only memorizing the noise pattern of one software tool. C. Hunting in the Shadows (Frequency Analysis) I notice that to break the overfitting cycle researchers look for features, in the frequency domain. Deepfakes are rarely made at resolution. Deepfakes are upsampled from a latent space. This upsampling often uses transposed convolutions. The upsampling adds a repeated grid-like flaw to the frequency spectrum. The repeated grid-like flaw appears as a checkerboard pattern. The checkerboard pattern cannot be seen by the eye. It is clear, in a Fourier Transform. Frank et al.. Others showed that analyzing anomalies can expose deepfakes regardless of the generator used [12]. EfficientNet [8] was capable, in this domain. EfficientNets ”compound scaling” architecture let EfficientNet balance resolution and depth. Let EfficientNet become a tool for excavating high-frequency noise fingerprints. I have seen the approach work in practice. Yet the approach is now, under threat. In my work I notice that modern diffusion models use noise scheduling and denoising processes. I notice that modern diffusion models do not leave the scars as older GANs [17]. I notice that diffusion models frequency based detectors. D. The Context Revolution (Transformers) This limitation brings us to a frontier: **Vision Transformers (ViT)**. CNNs act like a microscope looking at a painting. International Journal of Research in Engineering & Science, ISSN:P-2572-4274, O-2572-4304 volume 9 Issue 6 2025 Available online on http://rspublication.com/IJRES/IJRE.html DOI:10.5281/zenodo.17935569 ____________________________________________________________________________________________________________________ ©2025 RS Publication, [email protected] _________________________________________________________________________________________________________________________________________________ 208 II. LITERATURE SURVEY Transformers act like stepping to see the gallery. I see that Coccomini et al. ViTApp and other researchers argue that the next generation of detectors must understand context, not texture. I have seen a deepfake that can show skin texture but the deepfake does not match the lighting, in the environment. I have also seen a shadow on the nose that can suggest a source from the left while the background can suggest a source, from the right. CNNs have a view so CNNs cannot connect those far apart points. Transformers use self-attention so Transformers can handle those connections. I see that the method can query the relationship, between the eye and the background at the time. I see that the trade-off is cost. The method is heavy, slow and data hungry. Table I shows the evolution. I see that the gap, in the literature is clear. The current literature moves from checks to pixel checks and now to checks. The logical next step is the focus of this paper. The logical next step means that we should stop choosing between these methods and instead combine them into one framework. I put my approach in context. I examine the mechanisms that prior art uses. A. Spatial and Frequency Analysis Early forensic methods were manual. The early forensic methods looked for artifacts. The early forensic methods also looked for color mismatches. When generation quality improved researchers turned to the frequency domain. The researchers used the frequency domain to find clues. GANgenerated images often show anomalies in high-frequency bands. The spectral anomalies come from upsampling operations such, as transposed convolutions. I have seen the upsampling operations leave traces. Techniques that use Discrete Cosine Transforms (DCT) or Fourier Transforms isolate the fingerprints [12]. To automate the process, CNNs like EfficientNet [8] used compound scaling to keep the network depth, width and resolution, in balance. CNNs, like EfficientNet [8] could then capture high-frequency noise patterns with ease. When I looked at the work of Chen et al. DiffusionDF I saw that diffusion models work differently from GANs. Diffusion models often leave traces. Older CNNs often miss those traces. B. Vision Transformers and Global Reasoning The Vision Transformer (ViT) [9] broke away, from the approach. The Vision Transformer (ViT) does not use convolution. In my work I see that the Vision Transformer (ViT) does not look at pixels in neighborhoods. The Vision Transformer (ViT) uses MultiHead Self-Attention (MSA) to compare every patch of an image at the time. The Vision Transformer (ViT) can then spot meaning errors, such, as a background that bends oddly around a face across the picture. Recent benchmarks [5] show that the Vision Transformer (ViT) handles image changes better. The Vision Transformer (ViT) often needs training data to finish training. C. Ensemble Learning Ensemble learning combines predictions from models. Ensemble learning is a technique to reduce variance. Prior work, in deepfake detection has mostly used ensembles, such, as stacking ResNets. I think a heterogeneous ensemble can do more. Our work proposes an ensemble. The heterogeneous ensemble uses the feature extraction of CNNs and the global reasoning of Transformers. We wrote this section to explain the framework. The experimental framework shows the dataset composition. The experimental framework shows the architectures we used. The experimental framework also shows the formulation of the strategy. A. Dataset Composition We used the Labeled Deepfake Image Collection (LDIC) dataset [18]. The Labeled Deepfake Image Collection (LDIC) dataset provides a stress test, for generalization. The Labeled Deepfake Image Collection (LDIC) dataset contains images from ten generation pipelines. •Total Size: 18,124 images (256×256 pixels). •Class Balance: 5,890 Real images vs. 12,234 Fake images. •Generators used: I use the Generators DeepFaceLab, StyleGAN (v1/v2) FaceSwap, NeuralTextures, Midjourney and Stable Diffusion. I notice this variety helps the model learn to spot artifacts, from a tool. I also notice this variety makes the model learn the features of media. B. Architectural Formulation 1) EfficientNet-B0 (The Convolutional Stream): We chose EfficientNet because EfficientNet extracts details, with little cost. EfficientNet uses Mobile Inverted Bottleneck Convolution blocks, called MBConv blocks. EfficientNet also uses the compound scaling method. If we define a layer, as Yi=Fi(Xi)EfficientNet optimizes the network by scaling the depth (α) scaling the width (β) and scaling the resolution (γ) to: depth :d=αϕ,width :w=βϕ,resolution :r=γϕ s.t. α·β2·γ2≈2, α ≥1, β ≥1, γ ≥1 (1) When I use the method the model captures high frequency noise patterns of deepfake upsampling. The model stays easy to compute. 2) Vision Transformer (ViT-Tiny): To catch inconsistencies we used ViT-Tiny. We wanted to catch inconsistencies. The input image xgets reshaped into a sequence of flattened 2D patches xp. The input embedding is defined as: z0= [xclass;x1 pE;x2 pE;...;xN pE] + Epos (2) I notice that the linear projection E serves as the linear projection. I notice that the learnable position embeddings Eposserveasthepositionembeddings.T heT ransformerencoderbuildsalternatinglayersoftheHeadedSelf− AttentionMSAandtheMLPblocks.Theglobalreceptivefieldgivesthemodeltheabilitytofindfeaturesthat′refarapartinspace.T hemodelcanfindfeaturesorlightinggradientsthatdonotmatchthebackground. International Journal of Research in Engineering & Science, ISSN:P-2572-4274, O-2572-4304 volume 9 Issue 6 2025 Available online on http://rspublication.com/IJRES/IJRE.html DOI:10.5281/zenodo.17935569 ____________________________________________________________________________________________________________________ ©2025 RS Publication, [email protected] _________________________________________________________________________________________________________________________________________________ 209 IV. METHODOLOGY III. RELATED WORK TABLE I: Comparative Analysis of Key Detection Methodologies Author & Year Methodology Key Contribution Strengths Limitations Li et al. (2018) [3] Physiological Signal Analysis Detected lack of eye blinking Interpretability Obsolete against modern generators Afchar et al. (2018) [10] MesoNet Mesoscopic noise analysis Lightweight Low accuracy on high-quality fakes R¨ ossler et al. (2019) [13] XceptionNet Depthwise Separable Conv Industry Benchmark Poor cross-dataset generalization Li et al. (2020) FaceXray Face X-ray Blending boundaries Generalizable features Fails, on images that are not blended Coccomini et al. (2022) [15] Vision Transformers Self-attention mechanisms Global context awareness Data-hungry; computationally expensive C. Proposed Hybrid Ensemble Framework I propose a Stacking Ensemble that combines the probability distributions of the base learners. Unlike voting mechanisms the Stacking Ensemble learns the mix of models based on the validation performance. I think the Stacking Ensemble can improve the results. I feed the outputs of EfficientNet (PEff ) ViT (PV iT ) and Xception (PXcp) into a Logistic Regression meta-classifier. The meta-learner optimizes the weights wto minimize the log-loss: L=−1 N N X i=1 [yilog(ˆyi) + (1 −yi) log(1 −ˆyi)] (3) where the final prediction ˆyiis: ˆyi=σ(w1PEff +w2PV iT +w3PXcp +b)(4) I notice that dynamic weighting lets the system choose the model that’s most confident, for each input sample. I notice that dynamic weighting also fills the spots that each individual architecture has so the system does not miss anything. Fig. 1: The Proposed Hybrid Ensemble Architecture, illustrating the fusion of CNN and Transformer feature streams. D. Training Configuration We trained the models with PyTorch on an NVIDIA RTX 2050 GPU. The class imbalance had a ratio of about 1:2 so we used a Weighted Random Sampler. We optimized the models, with Adam set the learning rate to 1e−4and added a Cosine Annealing scheduler to fine-tune the weights as they approached convergence. Table II shows the performance metrics. The quantitative performance metrics cover all evaluated models, on the test set. TABLE II: Performance Comparison of Evaluated Architectures Model Accuracy Precision Recall F1-Score AUC Baseline CNN 0.3251 0.0000 0.0000 0.0000 0.4696 Xception-41 0.3376 0.9048 0.0207 0.0405 0.5026 ViT-Tiny 0.9945 0.9967 0.9951 0.9959 0.9995 EfficientNet-B0 0.9982 0.9995 0.9978 0.9986 1.0000 Ensemble 0.9993 1.0000 0.9989 0.9995 1.0000 A. Analysis of Results I look at the confusion matrices (Figures 2–6). I notice the performance differences clearly. •Baseline CNN & Xception: I saw that Baseline CNN & Xception did not learn all. Baseline CNN & Xception got accuracies, around 33 percent, which’s worse, than guessing in a balanced case. Baseline CNN & Xception could not handle the loss surface of the LDIC dataset, which has sources. Xception depended on DeepFaceLab artifacts that appear in FaceForensics++. Xception failed when Xception faced Diffusion-based images. •EfficientNet-B0: EfficientNet-B0 showed separation. EfficientNet-B0 identified artifacts, with precision. EfficientNet-B0s compound scaling let EfficientNet-B0 detect the noise patterns of the models. •Ensemble: The ensemble got scores. When we combined the predictions the ensemble filtered out the positives left from the single models. The ensemble achieved 100% precision and 99.89% recall. International Journal of Research in Engineering & Science, ISSN:P-2572-4274, O-2572-4304 volume 9 Issue 6 2025 Available online on http://rspublication.com/IJRES/IJRE.html DOI:10.5281/zenodo.17935569 ____________________________________________________________________________________________________________________ ©2025 RS Publication, [email protected] _________________________________________________________________________________________________________________________________________________ 210 V. EXPERIMENTAL RESULTS (a) Confusion Matrix (b) ROC Curve Fig. 2: Baseline CNN Performance: I see that the model does not separate the classes well. (a) Confusion Matrix (b) ROC Curve Fig. 3: Xception-41 Performance: Showing severe overfitting and poor recall. (a) Confusion Matrix (b) ROC Curve Fig. 4: ViT-Tiny Performance: Strong global context modeling. (a) Confusion Matrix (b) ROC Curve Fig. 5: EfficientNet-B0 Performance: Highly accurate texture analysis. (a) Confusion Matrix (b) ROC Curve Fig. 6: Proposed Ensemble Performance: The most robust predictor across all metrics. The results prove the idea. The heterogeneous ensembles beat the architectures because the heterogeneous ensembles cover spots. I see that EfficientNet finds frequency noise such, as upsampling artifacts, in GANs. I see that ViT finds structure problems that often appear in Diffusion models. I think the failure of Xception-41 is especially notable. Xception-41 was the model, on the FaceForensics++ dataset, which’s largely GAN based. Xception-41s separable convolutions seem to lack the ability to model the artifacts of diffusion models in LDIC. The failure of Xception-41 shows the need to move beyond 2019 era designs, for tasks. The failure of Xception-41 also shows the danger of relying on tests that do not show the variety of tools. Even though the performance looks good we see limitations, in the approach. The limitations are: 1) Temporal Blindness: In my experience the system looks at pictures. Temporal Blindness cannot see video problems such, as flickering over time or lip sync errors. Those problems need 3D CNNs or LSTM based networks. 2) Compression Robustness: Real world media, on the platforms like WhatsApp or TikTok get compression. We have not yet tested Compression Robustness for EfficientNets analysis, against JPEG artifacts. International Journal of Research in Engineering & Science, ISSN:P-2572-4274, O-2572-4304 volume 9 Issue 6 2025 Available online on http://rspublication.com/IJRES/IJRE.html DOI:10.5281/zenodo.17935569 ____________________________________________________________________________________________________________________ ©2025 RS Publication, [email protected] _________________________________________________________________________________________________________________________________________________ 211 VI. DISCUSSION VII. LIMITATIONS 3) Vulnerability: We have not tested the models, against attacks (FGSM, PGD). In those attacks the noise is added on purpose to fool the detector. 4) Generalization: Cross-Dataset Generalization matters. Even though LDIC is diverse Cross-Dataset Generalization needs testing, on datasets to see if the model overfits to a domain. The research aims to cut down misinformation. The research knows that detection research has a ”use” nature. The research shows model spots for example the failure of Xception. Bad actors could use those insights to build to detect deepfakes. The research thinks that being open, about defense tools is key, to building an information system. The research did not train or release any models. The direct risk of misuse stays low. I ran a study of the deep learning models, on the multi-source deepfake dataset. I saw that the traditional CNNs have trouble with the kinds of the generative models. I also saw that the new deep learning models such, as EfficientNet-B0 and ViT-Tiny give detection. We saw that the hybrid stacking ensemble got the scores, on every metric. The stacking ensemble proves that mixing the eye of convolutions with the global brain of transformers creates a very good method, for digital forensics. Future work will focus on the steps: •Temporal Deepfake Detection: Incorporating TimeSformer models to capture motion anomalies in video. •Adversarial Robustness: Evaluating and hardening the models against perturbation attacks. •Mobile Deployment: Mobile Deployment creates small versions of the ensemble. Mobile Deployment runs in time on applications. •Explainability: Integrating XAI methods (like GradCAM) to provide visual evidence of manipulation to users. REFERENCES [1] T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive Growing of GANs for Improved Quality, Stability, and Variation,” ICLR, 2018. [2] J. Chesney and D. Citron, “Deepfakes and the Threat to Truth,” California Law Review, 2019. [3] Y. Li, M.-C. Chang, and S. Lyu, “In Ictu Oculi: Exposing AI Created Fake Videos by Detecting Eye Blinking,” IEEE AVSS, 2018. [4] X. Yang, Y. Li, and S. Lyu, “Exposing Deep Fakes Using Inconsistent Head Poses,” ICASSP, 2019. [5] X. Wang et al., “DeepFakeBench: A Comprehensive Benchmark of Deepfake Detection,” NeurIPS, 2023. [6] A. Haliassos, J. Vougioukas, S. Zafeiriou, “Leveraging Real Talking Faces via Self-Supervision for Robust Forgery Detection,” ICCV, 2021. [7] F. Chollet, “Xception: Deep Learning with Depthwise Separable Convolutions,” CVPR, 2017. [8] M. Tan and Q. Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” ICML, 2019. [9] A. Dosovitskiy et al., “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,” ICLR, 2021. [10] D. Afchar et al., “MesoNet: A Compact Facial Video Forgery Detection Network,” WIFS, 2018. [11] H. Zhou et al., “Learning Rich Features for Deepfake Detection,” ACM MM, 2022. [12] Q. Yang et al., “Exposing Deepfakes via Multi-Attentional Frequency Analysis,” CVPR, 2021. [13] . Colleagues, ”FaceForensics++: Learning to Detect Manipulated Facial Images ” the ICCV, 2019. [14] L. Li et al. ”The Face X-ray technique detects face forgeries ” CVPR 2020. [15] D. Coccomini et al., “Combining EfficientNet and Vision Transformers for Video Deepfake Detection,” ICIP, 2022. [16] T. Nguyen et al., “Deepfake Generation and Detection: A Survey,” ACM Computing Surveys, 2023. [17] L. Chen et al., “On the Detection of Diffusion-Generated Deepfakes,” arXiv:2306.00045, 2023. [18] Jayanth Bottu, the Labeled Deepfake Image Collection, Kaggle, 2024. International Journal of Research in Engineering & Science, ISSN:P-2572-4274, O-2572-4304 volume 9 Issue 6 2025 Available online on http://rspublication.com/IJRES/IJRE.html DOI:10.5281/zenodo.17935569 ____________________________________________________________________________________________________________________ ©2025 RS Publication, [email protected] _________________________________________________________________________________________________________________________________________________ 212 VIII. ETHICAL CONSIDERATIONS IX. CONCLUSION AND FUTURE WORK